Friday, April 20, 2018

MySQL adjustment bureau

When maintainng any piece of software, we usually deal with two kind of actions:

  • bug fixing,
  • new features.

bugs and features

A bug happens when there is an error in the software, which does not behave according to the documentation or the specifications. In short, it's a breech of contract between the software maintainer and the users. The promise, i.e. the software API that was published at every major version, is broken, and the software must be reconciled with the expectations and fixed, so that it behaves again as the documentation says. When we fix a bug in this way, we increment the revision number of the software version (e.g. 1.0.0 to 1.0.1. See semantic versioning).

New features, in turn, can be of two types:

  • backward compatible enhancements, which add value to the software without breaking the existing functionality. This is the kind of change that requires an increment of the minor indicator in the version (for example: 1.1.15 to 1.2.0.)
  • Incompatible changes that break the existing behavior and require users to change their workflow. This kind of change requires bumping up the major number in the version (as in 2.1.3 to 3.0.0.)

Not a bug, nor a feature, but an adjustment.

The above concepts seem simple enough: you either fix something that's broken or add new functionality.

However, when maintaining a tool that has the purpose of helping users to deal with another software (as it is the case of dbdeployer that helps users to deploy MySQL databases) there is yet another category of changes that don't fall into the standard categories: it's what happens when the software being helped (MySQL) changes its behavior, which would break the normal functioning of the helping tool, giving the maintainer a difficult choice:

  • shall I modify the tool's interface to adapt to the new behavior, breaking existing procedures?
  • or shall I adapt the tool's functioning behind the scenes to keep the interface unchanged?

My philosophy with dbdeployer (and MySQL-Sandbox before it) is to preserve the tool's interface, so that users don't have to change existing procedures. I call this kind of changes adjustments, because they are not bugs, as they are not a consequence of a coding error, and not a feature, as the intervention is not a conscious decision to add new functionality, but an emergency operation to preserve the status quo. You can think of this category as a capricious change in specifications, which so often happens to software developers, with the difference that the one changing the specs is not the user, but a third party who doesn't know, or care, about our goal of preserving the API integrity.

For example, from MySQL 8.0.3 to 8.0.4 there was a change in the default authentication plugin. Instead of mysql_native_password, MySQL 8.0.4 uses caching_sha2_password. The immediate side effect for MySQL-Sandbox and dbdeployer was that replication doesn't work out of the box. A possible solution would be to force the old authentication plugin, but this would not allow users to test the new one. Since the main reason to use a tool like dbdeployer is to experiment with new releases safely, I had to keep the default behavior. Thus, I left the default plugin in place, and changed the way the replication works. It's an ugly workaround actually, but allows users to see the new behavior without losing existing functionality.
To complete the adjustment, I added a new option --native-auth-plugin, which would deploy using the old mysql_native_password. In total, the adjustment consists of a behind-the-scenes change, almost undetectable by users, and a new option to keep using the familiar authentication if users want it.

From the point of view of semantic versioning, this kind of change is a backward-compatible modification of the API, which warrants an increase of the minor number of the version.

Another example: when MySQL went from 8.0.4 to 8.0.11, it introduced a deal breaker change: the X Plugin is now loaded by default. This is easy for users of MySQL as a document store, as they don't need to enable the plugin manually, but bad news for anyone else, as the server is opening a port and a socket that many users may not choose to open voluntarily. What's worse, when installing more sandboxes of version 8.0.11 in the same host (for example in replication), one will succeed in reserving the plugin port and socket, while the others will have the error log populated with surprising errors about a socket being already in use.

The solution is similar to the previous one. When dbdeployer detect MySQL 8.0.11 or newer, it adds options to customize the mysqlx plugin port and socket, thus allowing a frictionless deployment where the new functionality is available to the brave experimenters. At the same time, I added a new option (--disable-mysqlx) for the ones who really don't want an extra port and socket in their servers, not even for testing.

These adjustment are usually costly additions. While the added code is not that much, they require extra tests, which are often complex and require more time to write and execute them. The process to add an adjustment goes mostly like this:

  • I dedicate my morning walk to think about the fix. Sometimes the fix requires several walks, while I decide the less intrusive solution.
  • If the walk has been fruitful, writing the code requires just a few minutes. If I missed something, I iterate.
  • Then the more difficult part: writing meaningful tests that prove that the adjustment is correct and it doesn't introduce side effects in any MySQL version. And of course the option that reintroduces the old behavior must be tested too.
  • A positive side effect of this exercise is that often I realize that I was missing a test for an important behavior and then I write down that as well. The test suite included 6,000+ tests 1 month ago, and now it has almost doubled.


Mark Callaghan said...

Your process sounds wonderful (go for a walk, think, repeat).

Giuseppe Maxia said...

It's a proven algorithm.

"Throughout history, great minds in literature, philosophy, and science have found important insight and inspiration while out on a walk. Perhaps this is because walking – at least while out in nature (which is the kind of walking many of these thinkers favored) – has been shown by modern science to improve memory and attention. Or perhaps it’s because walking simply gets the blood pumping – a hard to quantify effect of invigoration."